On Cohesively Polarized Communities in Signed Networks

Locating and characterizing polarization is one of the most important issues to enable a healthier web ecosystem. Finding groups of nodes that form strongly stable agreements and participate in collective conflicts with other groups is an important problem in this context. Previous works approach this problem by finding balanced subgraphs, in which the polarity measure is optimized, that result in large subgraphs without a clear notion of agreement or conflict. In real-world signed networks, balanced subgraphs are often not polarized as in the case of a subgraph with only positive edges. To remedy this issue, we leverage the notion of cohesion — we find pairs of cohesively polarized communities where each node in a community is positively connected to nodes in the same community and negatively connected to nodes in the other community. To capture the cohesion along with the polarization, we define a new measure, dichotomy. We leverage the balanced triangles, which model the cohesion and polarization at the same time, to design a heuristic that results in good seedbeds for polarized communities in real-world signed networks. Then, we introduce the electron decomposition which finds cohesively polarized communities with high dichotomy score. In an extensive experimental evaluation, we show that our method finds cohesively polarized communities and outperforms the state-of-the-art methods with respect to several measures. Moreover, our algorithm is more efficient than the existing methods and practical for large-scale networks.